DEVELOPMENT OF SUPERCOMPUTER IMAGE PROCESSING SOFTWARE WITH X-WINDOW
USER-INTERFACE FOR THE PROCESSING OF THE REMOTELY SENSED DATA
Young-Kyu Yang,
Hyun-Ok Nam,
Kyoung-Ok Kim,
Seong-Ik Cho
Systems Engineering Research Institute
Korea Institute of Science and Technology
1 Eoeun-dong, Yoosung-ku
Taejeon, Korea
305-333
COMMISSION II
ABSTRACT
This paper presents the design concept, the functions, and the examples of utilization of C-ERIMS
(Environment and Resources Integrated Management System for CRA Y) image processin~ software
package under development on Cray-2 supercomputer system. C-ERIMS is equipped wlth almost
all the commonly used image processing algorithms as well as special features like parallel
processing, neural network, 3-D perspective animation, etc. It has interactive true color display
capability utilizing x-window on the CRA Y supercomputer system under UNICOS operating system.
KEY
WORDS:
Image Processing,
l.
INTRODUCTION
Supercomputer,
This software is code named C-ERIMS which
is an acronym for Environment and Resources
Integrated Management System for CRA Y. CERIMS is designed to operate as a complete
user-friendly interactive system with integrated
software and open Graphic User Interface
(GUI), i.e. X Window system.
2.2
C-ERIMS system can:
The philosophy behind the developing C-ERIMS
was to take advantage of the computing power
and large main memory of supercomputer and
provide
the
resource
mangers
and
other
decision makers with a powerful tool necessary
for their national scale resource management.
X Window System
The architecture of the X Window system is
based on a client-server model. The X server
controls all resources which include windows,
bitmaps, fonts, colors, and other data structures
used by the window system to enable clients
to
use
and
share
these
data
structures
transparently. Network transparency implies that
applications can run on whatever CPU is most
convenient. For example, applications requiring
extensive computations ,can run on a networkconnected supercomputer and the result can be
displyed on the terminal or workstation.
BASIC DESIGN CONCEPT
Software Design
A complete survey
processing
software
X-window
The X Window system developed at the MIT
(Massachusetts Institute of Technology) seems
to satisfy the needs of users for highperformance, high functionality, network based
system for high-resolution graphics. The purpose
of the X Window system is to provide a
network-transparent
and
vendor-independent
operating environment for workstation software
and to maintain portability with a common
programming interface across multiple vendor
platforms.
o directly access satellite images and other
remotely sensed data,
o extract subsections or full scene to disk
files,
o act upon these files to perform the
generally available image processing
functions, and
o generate hard copies of processed images.
2.1
Parallel Processing,
maximum efficient design of the C-ERIMS
system. The CSADIE (Samayoa, 1983) and
GIPSY (Gu and Samayoa, 1988) software
packages available on Cray computer system
seem to have limited capability due to their
batch oriented operation and lack of true color
manipulation capability. They were originally
developed on smaller computer system such as
workstation or Cyber system and later ported
to the Cray computer. Other systems including
VICAR,
ELAS
(NASAjERL,
1988),
and
LARSYS (Williams, et at, 1976) were also
surveyed and analyzed. A comprehensive image
processing software was successfully designed to
handle full scene satellite imagery in interactive
way using X Window protocol.
A powerful supercomputer based remote sensing
image processing software package is being
developed at the Korea Institute of Science
and Technology (KIST). KIST have experience
of developing image processing software on
microcomputer
and
main
frame
computers
(Yang, et aI., 1989a; Yang, et aI., 1989b).
2.
Neural Network,
for the existing image
was
performed
for
235
3.
3.1
SUPERCOMPUTING
4.
ENVIRONMENT
PROCESSING
MODULES
Most of commonly available image procesing
functions are being implemented. It includes
preprocessing,
image
enhancement,
image
transformation, image classification, overlaying
and mosaicking, data compression, and some
utility functions.
Cray-2S System
Cray-2s/4-128 supercomputer was used as the
platform for software development. It has four
CPU s and equipped with 128 mega word
SRAM main memory and 40 giga bytes of
high speed disk. The clock cycle is 4.1 nano
second and its peak performance is 2 GFLOPS
(Giga Floating Point Operations per Second).
The operating system is UNICOS system which
is based on AT & T UNIX System V and
enhanced to support both batch and interactive
processing in a large scale scientific computer
environment. UNICOS supports CAL, the Cray
Macro Assembler, and the high level languages
including Fortran, C, and Pascal. The Fortran
compiler
has
capability
of
automatically
vectorizing inner DO-loops, and thus provides
program optimization functionalty.
3.2
IMAGE
4.1
Preprocessing
The sources of geometric error are instrument
error, panoramic distortion, earth rotation and
platform instability. The geometric correction
process
comprises
the
determination
of a
relationship between the coordinate system of
map and image. It establishes a set of points
defining pixel centers in the corrected image
that, when considered as a rectangular grid,
defines an image with the desired cartographic
properties, and the estimation of pixel values
to be associated with those points. The cubic,
bilinear, and nearest neighborhood method are
available as the resampling methods.
Network
The Cray-2S system is connected to IBM 3083,
and MicroVax system through 100 mega bytes/
second HYPER Channel (Figure 1). Local
Ethernet network is also connected to Cray
system through TCP/IP networking protocol. A
remotely located NAS AS/XL V50 and a CDC
Cyber
960-31
mainframe
computers
are
connected to Cray system via dual 1. 5 mega
bits per second T1 communication line. The
Unix
workstations
including
Sun,
Silicon
Graphic's IRIS are connected to Ethernet
network
and
thus
provides
users
with
accessibility to any of the existing computer
resources including Cray, NAS, Cyber, Vax,
and other workstations.
4.2 Image Enhancement
Image
enhancement
is
to
enhance
the
weakened image and to increase visible effect.
Contrast stretch method (linear, normalized,
histogram
equalization,
exponential,
and
logarithmic methods) for image enhancement
were implemented. Spatial filtering with mean,
mode, Roberts gradient, difference operator,
and sobel operator were also implemented. FFT
(Fast Fourier Transformation), contouring, and
color enhancement is also available.
4.3
Image transformation
Various
image
transformation
algorithm
is
available. They can be summarized as follows:
1)
Arithmetic operation:
o band ratio
o band difference
2)
Vegetation index
o normalized difference
o leaf area index
o transformed vegetation index
o tasselled cap transformation
o perpendicular vegetation index
3)
Principal Component Analysis
4.4 Classification and
1)
2)
Fig.
1.
Supercomputing environment
in KIST.
236
o sum
o add
Other Algorithms
Classification
o supervised
Terrain analysis
o slope-gradient
o shaded relief
o unsupervised
o slope-aspect
3)
Three dimensional perspective view
(Figure 2)
o coordinate transformations
o projection transformation
o hidden line procedure
o scaling
4)
Generation of mosaic map (Figure 3)
o density level normalization
o correction of the density level by
overlapped area
o seam control point matching
5)
Utility routines
o data file handling
- file format conversion
- band extraction
o coordinate conversion
- UTM versus geographic
coordinate
o color map manipulation
o histogram computation
o statistical analysis
o image compression
4.5 Time
5.
5.1
NEURAL NET ALGORITHM
Introduction
Recently efforts to adopt neural net algorithm
to sa tellite imagery analysis have been very
active (Benediktsson, et al., 1990; Heermann
and Khazenie, 1992; Lee, et aI., 1990; Ryan,
et al. , 1991) .
Comparison with Other Systems
Test run of the several algorithms on Cray-2S
computer
shows
more
than
50
times
improvement in processing speed comparing to
the
IBM
3083
and
VAX
system.
The
algorithms used for test are as follows:
o G CP correction
o 3-D perspective view display
o Maximum likelihood classification
A neural net algorithm and software named
"atree (Adaptive TREE)" was adopted and
implemented in Cray-2S.
It was originally
developed
for
PC
and
workstations
by
University of Alberta, Canada (Dwelly, 1990).
The atree contains the learning algorithm for
adaptive
logic
netwoks,
which
is
the
backpropagation algorithm for multilayer feed
forward artificial neural networks. The atree
algorithm demostrates that for those networks
which is consisting of multilayer threshold,its
elements can be effectively trained.
It is an important property for logic networks
to have ability to generalize their responces to
new
inputs,
presented
after
training
is
completed.
The
successful
generalization
properties of these logic networks are based
on the observation, backed up by a theory,
that trees of "two-input" logic gates of types
AND, OR, LEFT and RIHGT are very
insensitive to changes of their inputs.
5.2
Fig.
2.
"A tree" Algorithm Concept
An atree (Adaptive TREE) is the fundamental
structure used by routines in this software
(Dwelly, 1990). This is a binary tree with
nodes of two types: (1) adaptive elements, and
(2) leaves. Each element takes one of four
logical
functions,
AND,
OR,
LEFT,
or
RIG HT depending on initialization or the
training it has undergone. The leaf nodes of
the tree serve only to collect the inputs to
the subtree of elements. Each one takes its
input bit from a boolean input vector. The
tree produces a single bit as its output. For
computing non-boolean outputs, several trees
are used in parallel to produce a vector of
bits representing the output value.
Three dimensional display of
Typhoon "A be".
5.3
"Atree" Implementation
The original routines written in C language by
University of Alberta to create, train, evaluate,
and print out adaptive logic networks were
transported to the Cray computer. The user
can create a training set, then create a tree
using 'atree_create'. The tree is trained using
'atree_train' and then it can be used to
evaluate new inputs using 'atree_evaI'. Because
a single tree produces only one bit, the
programmer must train several trees on the
input data, each one responsible for one bit
of the output data. This is made slightly
simpler by the choice of parameters for 'atree_
train' which takes an array of bit vectors as
the training set, and an array of bit vectors
for the result set. The figure 4 shows its
typical description.
Fig.
3.
6.
Mosaic image of middle part
of Korea.
6.1
USER
INTERFACE
Graphic User Interface
X-lib was selected as the graphics tool for C-
237
ERIMS system. It defines -an-extensive set of
functions and provides complete access and
control over the display, windows, and input
devices. It seems to become one of the most
widely used low-level graphic interface in Unix
systems.
Motif was selected as the user interface tool.
Many software developers prefer to use one
of the higher-level X toolkits since using lowlevel graphics library to develop application
software could be tedious and time consuming.
A
toolkit
should
contain
a
user-interface
subroutine library to simplify the design and
development
of application
user
interfaces.
Motif provides special tool kit feature named
widgets which supports human interface graphic
objects, such as dialogue box, menu, or mousesensitive button.
A minimal X application first sets up a
connection to the workstation, then creates a
window, creates X resources, maps the window
to the screen to make it visible, solicits the
input events it is interested in, reads and
interprets
these
input
events,
and
finally
generates graphical output.
Compilation of an X application program
in Cray computer:
-IX 11
xhost Cray2s
telnet Cray2s
userid: xxx
passwd: yyy
setenv DISPLAY
6.3
irisOOO
X Toolkit Programming
In X Toolkit programming, Xt Intrinsics serves
as a framework that allows programmers to
create
user
interfaces
by
combining
an
extensible set of user interface components.
These components are known as widgets, and
include menus,
dialogue boxes,
title bars,
scrollbars, selection boxes, labels and push
buttons.
performs
several
basic
o Initialize the Intrinsics.
X tlnitializeO esta bUshes
a connection to the X server.
o Create widgets.
XtCreateWidget(name, class,
parent, args, nargs)
o Register callbacks and event
handlers.
Xt
o Realize all widgets.
o Enter the event loop.
In order to run an X Window based software,
following procedure is necessary.
cc filename
Designating the display device:
X Toolkit application
steps. These are:
6.2 Execution of an X program
1)
2)
The compilation step would be:
-lnet
cc -0 filename 1 filename2
-X 11
y
7.
-IXm
-iXt
PARALLEL PROCESSING
In order to benefit from the power of vector
hardware, existing programs should be rewritten
or translated to the parallel form. The existing
auto vectorizing compiler does not provide
maximum efficient method to discover the
parallelism implicit in a Fortran program.
An efficient translation algorithm is being
developed as shown in figure 5. The scannerparser phase converts the input program to an
abstract syntax tree that is used as the
intermediate form throughout the translation.
The pretty printer can reconstruct a source
program from the abstract syntax tree; it is
used throughout the translator. The vector
translation
phase
consists
of
three
main
subphases:
Trees -- one
per output bit
Random
connections
'--"-"l...--'
(1)
Complements
(2)
(3)
Xl
subscript standardization,
which encompasses all the
transformation that attempt to
put subscripts into canonical
form;
dependence analysis, which
builds the interstatement
dependence graph;
parallel code
generation,
which
generates
array
assignments where possible.
X2
8.
Fig.
4.
Description of A tree structure (Dwelly, 1990).
CONCLUSION
A Cray-2s based image processing system code
named C-ERIMS is being developed. It will
238
have full scene handlirig capability tak~g
advantage of high speed and large mam
memory of the supercomputer. It's software will
be composed of the most of the commonly
available
functions
including
preprocessing,
image
enhancement,
image
transformation,
classification,
image
compression,
three
dimensional perspective view generation, mosaic
map generation and terrain information analysis.
Most of the image processing modules are
written in Fortran and C. Test run of the
some algorithm in C-ERIMS on the Cray
showed more than 50 times improvement in
processing speed comparing to the IBM and
V A X systems.
Data Using a Back-Progation Neural Network,
IEEE Transactions on Geoscience and Remote
Sensing, 30(1):81-88.
Dwelly, Andrew,
Adaptive
Logic
Alberta, Canada.
1990. An Implementation of
Networks,
University
of
25 pg.
Gu, Goson and B. Samayoa, 1988. Conversion
of Gipsy Code on Cray Unicos System, The
Final Report on 1988 Summer Intern Project
in Digital Image Processing, Cray Research
Inc., 31 pg.
Lee, J, R. C. Weger, S. K. Sengupta and
R.
M.
Welch,
1990. A Neural Network
A pproach
to
Cloud
Classification,
IEEE
Transactions
on
Geoscience
and
Remote
Sensing, 28(5) :846-855.
C-ERIMS has user friendly interactive capability
on X-terminal or UNIX workstations connected
to Cray system via X Window protocol. The
X lib was adopted as the graphic tool and
Motif tool kit
provides easy-to-use
menu
system. It will also be equipped with special
features such as neural net algorithm, parallel
processing algorithm, and some of knowedge
based inference algorithm
to enhance the
processing
capability
and
the
classification
accuaracy. C-ERIMS system is expected to be
available to the pubhc by the end of 2nd
quarter next year (1993).
1988.
ELAS: Earth Resources
Laboratory Application Software. NASA/ERL
NSTL Report 183. NSTL, MS.
NASA/ERL,
Ryan W., P. J. Sementilli, P. Yuen, and B.
R.
Hunt.
1991.
Extraction
of
Shoreline
Features by Neural Nets and Image Processing.
Photogrammetric
Engineering
and
Remote
Sensing, 57(7) :947-955
Samayoa, William,
1983. CSADIE 3.0 for
Cray Computers: System Documentation. Cray
Research Inc., Mendota Height, Mn. 175 pg.
tree
Scanner
Parser
tree
Vector
Translator
Preliminary
Transforms
(standardization,
dependence analysis)
Fig.
5.
9.
Williams, G. N., et al., 1976. Conversion of
LARSYS III. 1 to an IBM 370 Computer.
Final Report, Contract NAS9-14514 by Texas
A&M Univ. for Earth Observation Division,
NASA Johnson Space Center, Houston, TX.
60 pg.
Pretty
Printer
Yang, Y. K., S. E. Cho, et aI., 1989a.
Development
of
a
Microcomputer
Image
Processing Systems for Analyzing Satellite and
Airborne
Sensor
Data.
MOST
National
Research Report BS N20622. 173 pg.
Parallel Code
Generation
Yang, Y. K., S. E. Cho, et al., 1989b.
Development
of
Satellite
Image
Processing
Software on Mainframe Computer. Journal of
Korean Society of Remote Sensing, 5(1):29-39.
Overview of Translation Process
ACKNOWLEDGEMENTS
The
authors
would
like
to
thank
Cray
Research,
Inc.
for partially supporting CERIMS
development
project
through
Cray
University Grant Program.
REFERENCES
Benediktsson, J. A., P. H. Swain, and O. K.
Ersoy,
1990.
Neural
Network
Approaches
Versus Statistical Methods in Classification of
Multisource
Remote
Sensing
Date,
IEEE
Transactions
on
Geoscience
and
Remote
Sensing, 28(4): 540-551.
Heermann, P. D. and N. Khazenie, 1992.
Classification of Multispectral Remote Sensing
239